Hi Steven! You are right, there could be some cascading effect from previous checkpoints. I think the best way to handle that is to set the "minimum pause between checkpoints". In fact, I would actually recommend this over the checkpoint interval parameter.
The pause will allow the job to handle such effects that built up during an unhealthy checkpoint. You can for example set the checkpoint interval to 2 mins and set the pause to 1.5 mins. That way, if a checkpoint takes longer than usual, the next one will still wait for 1.5 mins after the previous one completed or expired, giving the job time to catch up. Best, Stephan On Fri, Dec 1, 2017 at 10:10 PM, Steven Wu <stevenz...@gmail.com> wrote: > Here is the checkpoint config. no concurrent checkpoints with 2 minute > checkpoint interval and timeout. > > Problem is gone after redeployment. I will try if I can reproduce the issue > > [image: Inline image 1] > > On Fri, Dec 1, 2017 at 6:17 AM, Nico Kruber <n...@data-artisans.com> > wrote: > >> Hi Steven, >> by default, checkpoints time out after 10 minutes if you haven't used >> CheckpointConfig#setCheckpointTimeout() to change this timeout. >> >> Depending on your checkpoint interval, and your number of concurrent >> checkpoints, there may already be some other checkpoint processes >> running while you are waiting for the first to finish. In that case, >> succeeding checkpoints may also fail with a timeout. However, they >> should definitely get back to normal once your sink has caught up with >> all buffered events. >> >> I included Stefan who may shed some more light onto it, but maybe you >> can help us identifying the problem by providing logs at DEBUG level >> (did akka report any connection loss and gated actors? or maybe some >> other error in there?) or even a minimal program to reproduce. >> >> >> Nico >> >> On 01/12/17 07:36, Steven Wu wrote: >> > >> > org.apache.flink.runtime.checkpoint.CheckpointCoordinator - Checkpoint >> > 9353 expired before completing >> > >> > I might know why this happened in the first place. Our sink operator >> > does synchronous HTTP post, which had a 15-mint latency spike when this >> > all started. This could block flink threads and prevent checkpoint from >> > completing in time. But I don't understand why checkpoint continued to >> > fail after HTTP post latency returned to normal. there seems to be some >> > lingering/cascading effect of previous failed checkpoints on future >> > checkpoints. Only after I redeploy/restart the job an hour later, >> > checkpoint starts to work again. >> > >> > Would appreciate any suggestions/insights! >> > >> > Thanks, >> > Steven >> >> >
